An environment compliance intelligent question and answer system based on causal diagram reasoning

Abstract

This invention discloses an intelligent environmental compliance question-answering system based on causal graph reasoning, comprising: a natural language input module for receiving natural language questions input by the user; a multimodal data acquisition module for collecting environmental inspection data; a multimodal feature encoding module for performing multimodal feature extraction processing; a causal graph construction module for constructing a hierarchical causal graph structure; a causal graph neural reasoning module for inputting the hierarchical causal graph structure into a hierarchical causal graph neural network; a causal state reasoning module for performing causal reasoning operations; a semantic matching and question-answering module for performing semantic matching and path retrieval; a causal chain visualization module for displaying the causal reasoning graph; and a compliance question-answering output module for outputting structured question-answering results and corresponding causal chain graphs. This invention employs a hierarchical causal graph neural network to achieve intelligent environmental compliance question answering and causal reasoning.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and environmental data processing technology, and in particular to an intelligent question-and-answer system for environmental compliance based on causal graph reasoning. Background Technology

[0002] With the increasing environmental management requirements of industrial enterprises, emission compliance testing and intelligent question-and-answer systems are gradually becoming an important direction for environmental information-based supervision. In existing technologies, environmental compliance analysis systems mostly rely on rule-based knowledge retrieval and text matching methods. This involves storing emission indicators, monitoring data, and regulatory texts in a database, and then performing keyword matching and conditional searches when users submit queries.

[0003] While these methods can achieve basic compliance Q&A functionality, they have significant limitations when dealing with complex, multi-source data. First, existing systems mostly use single-modal information as input and cannot simultaneously process multiple types of inspection data such as images, numerical values, and text, resulting in insufficient accuracy of compliance judgment results.

[0004] Secondly, traditional question-answering models often rely on shallow semantic similarity for matching, which fails to capture the deep causal relationships between emission behavior, violation status, and regulatory provisions, and lacks the ability to reason about the causes of violations and the logic that triggers regulations.

[0005] In addition, existing environmental intelligent question answering systems mostly adopt static knowledge graph structures with fixed node relationships, making it difficult to dynamically update causal dependencies based on actual inspection data, thus failing to achieve interpretable reasoning output.

[0006] In summary, existing technologies still have shortcomings in multimodal data fusion, causal relationship modeling, and interpretable question-answering reasoning.

[0007] Therefore, how to provide an intelligent question-answering system for environmental compliance based on causal graph reasoning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0008] One objective of this invention is to propose an intelligent question-and-answer system for environmental compliance based on causal graph reasoning. This invention utilizes multimodal feature extraction, hierarchical causal graph structure modeling, and hierarchical causal graph neural network reasoning techniques to fuse and analyze image, numerical, and textual inspection data of enterprise emission facilities. The system first extracts semantic features from different types of data through a multimodal feature encoding module, and then constructs a hierarchical causal graph structure containing data, behavior, and rule layers. Subsequently, a hierarchical causal graph neural network is used to perform causal propagation and node state updates, establishing a directional information transmission mechanism between layers to achieve causal dependency modeling between violation status, risk level, and regulatory clauses. Finally, a semantic matching mechanism is combined to complete question-and-answer generation and visualize the causal chain.

[0009] An environmental compliance intelligent question-answering system based on causal graph reasoning according to an embodiment of the present invention includes:

[0010] The natural language input module is used to receive natural language questions input by the user;

[0011] The multimodal data acquisition module is used to collect environmental inspection data of the emission facilities to be inspected, and obtain the raw inspection dataset;

[0012] The multimodal feature encoding module performs multimodal feature extraction processing on the original inspection dataset to generate multimodal causal node representations;

[0013] The causal graph construction module is used to construct a hierarchical causal graph structure based on the multimodal causal node representation and to make cross-layer connections through causal dependency edges;

[0014] The causal graph neural reasoning module is used to input the hierarchical causal graph structure into the hierarchical causal graph neural network, perform node state updates and causal propagation based on the directional attention mechanism, and generate a causal propagation result graph.

[0015] The causal state reasoning module is used to perform causal reasoning operations based on the causal propagation result diagram, determine the non-compliance status, risk level and triggering clauses of the target facility, and generate a causal reasoning result set;

[0016] The semantic matching and question answering module is used to semantically encode natural language questions and perform semantic matching and path retrieval with the causal inference result set to generate structured answer content.

[0017] The causal chain visualization module is used to display the causal reasoning graph based on the causal paths obtained through matching;

[0018] The compliance question-and-answer output module is used to output structured question-and-answer results and corresponding causal chain graphs.

[0019] Optionally, modules can be integrated using the following methods:

[0020] Collect environmental inspection data of the emission facilities to be tested, obtain additional attribute information associated with the inspection task, and obtain the original inspection dataset;

[0021] Multimodal feature extraction is performed on the original inspection dataset to obtain multimodal causal node representations;

[0022] Based on multimodal causal node representation, a data layer subgraph, a behavior layer subgraph, and a rule layer subgraph are established. Cross-layer directed connections are established between the subgraphs according to causal dependencies to construct a hierarchical causal graph structure.

[0023] The hierarchical causal graph structure is input into the hierarchical causal graph neural network, and causal propagation and node state update are performed to obtain the causal propagation result graph.

[0024] Based on the causal propagation result graph, perform causal reasoning operations to determine the non-compliance status, risk level, and triggering clauses of the target facility, and generate a causal reasoning result set;

[0025] It receives natural language questions input by users, performs semantic encoding on the natural language questions, performs semantic matching and path retrieval with the causal inference result set, and generates structured answer content;

[0026] It outputs structured answers and generates visual causal explanations based on causal chain paths, showcasing the complete reasoning path.

[0027] Optionally, the step of collecting environmental inspection data of the emission facilities to be inspected, obtaining additional attribute information associated with the inspection task, and obtaining the original inspection dataset specifically includes:

[0028] Environmental inspection data of the target emission facility is collected by scanning terminals or mobile inspection equipment configured at the emission facility site. The environmental inspection data includes image data, numerical detection data and text description data.

[0029] The image data consists of images of the facility's appearance, operating status, and emission outlet taken by the inspection personnel. The numerical detection data includes temperature, pressure, flow rate, pH value, COD, ammonia nitrogen concentration, and suspended solids content. The text description data includes operating status descriptions, maintenance records, and fault notes filled in by the inspection personnel.

[0030] Obtain additional attribute information associated with the inspection task, including facility type identifier, inspection timestamp, geographic location information, and responsible person number;

[0031] Image data, numerical detection data, text description data, and attribute information are uniformly encapsulated to generate the original inspection dataset.

[0032] Optionally, the step of performing multimodal feature extraction processing on the original inspection dataset to obtain multimodal causal node representations specifically includes:

[0033] The image data from the original inspection dataset is input into a multi-scale convolutional coding network to extract multi-scale features from the image information, thereby obtaining image feature vectors.

[0034] Numerical detection data is input into a numerical coding network, and each detection parameter is embedded, mapped, and transformed nonlinearly to obtain a numerical feature vector.

[0035] Input the text description data into the pre-trained language model encoder, perform contextual semantic modeling on the text content, and output the text feature vector;

[0036] Image feature vectors, numerical feature vectors, and text feature vectors are used as inputs, and a multimodal gating fusion module is used to perform fusion processing to obtain a multimodal causal node representation.

[0037] Optionally, the step of establishing data layer subgraphs, behavior layer subgraphs, and rule layer subgraphs based on multimodal causal node representation, and establishing cross-layer directed connections between each subgraph according to causal dependencies to construct a hierarchical causal graph structure specifically includes:

[0038] A data layer subgraph is constructed based on the multimodal causal node representation vector. A data layer node set is defined based on the feature variables related to emission indicators, facility status, inspection time and geographical location in the multimodal causal node representation. The data layer node set includes flow rate nodes, temperature nodes, pH value nodes, chemical oxygen demand nodes, facility type nodes, geographical location nodes and time nodes.

[0039] Assign corresponding node feature vectors to each data layer node, and use linear transformation weight matrix and bias term to map the multimodal causal node representation to obtain the feature representation of each data layer node;

[0040] A set of directed edges in the data layer is established based on the statistical dependencies and structural priors between data layer nodes. The directed edges in the data layer represent the causal directional relationships between nodes and constitute the set of data layer edges.

[0041] A data layer adjacency matrix is ​​generated based on the data layer edge set to represent the directed connection relationships between data layer nodes, forming a data layer subgraph containing node feature sets and adjacency relationships;

[0042] Construct a behavior layer subgraph and define a behavior layer node set based on the feature variables related to corporate emission compliance behavior in the multimodal causal node representation vector. The behavior layer node set includes violation judgment nodes, risk classification nodes, and exceedance level nodes.

[0043] Assign a corresponding node feature vector to each behavior layer node, and use a linear transformation weight matrix and bias term to map the multimodal causal node representation to obtain the feature representation of each behavior layer node;

[0044] The set of directed edges between behavioral layer nodes is determined based on historical violation records and risk assessment data. These directed edges represent the causal dependencies between violation status, risk level, and exceedance level.

[0045] A behavior layer adjacency matrix is ​​generated based on the behavior layer edge set to describe the directed connection relationships between behavior layer nodes, forming a behavior layer subgraph containing the node feature set and adjacency relationships;

[0046] Construct a rule layer subgraph and define a set of rule layer nodes based on the feature variables related to environmental regulations knowledge in the multimodal causal node representation vector. The rule layer nodes include regulatory clause nodes, penalty standard nodes, and rectification suggestion nodes.

[0047] Assign corresponding node feature vectors to each rule layer node, and use linear transformation weight matrix and bias term to map multimodal causal node representation vectors to obtain feature representation of each rule layer node;

[0048] A set of directed edges at the rule layer is established based on the citation relationships between legal provisions, the relationship between penalty standards and the application of provisions, and the relationship between rectification suggestions and the constraints of provisions. The directed edges at the rule layer are used to describe the causal dependencies between legal provisions, penalty standards, and rectification suggestions.

[0049] A rule-layer adjacency matrix is ​​generated based on the rule-layer edge set to represent the directed connection structure between rule-layer nodes, forming a rule-layer subgraph containing the node feature set and adjacency relationship;

[0050] Establish a set of cross-layer directed edges. Based on the semantic relationships between nodes in the data layer subgraph, behavior layer subgraph, and rule layer subgraph, establish sets of cross-layer directed edges between the data layer and behavior layer, and between the behavior layer and rule layer, respectively.

[0051] The cross-layer connection edges include directed edges from data layer nodes to behavior layer nodes and directed edges from behavior layer nodes to rule layer nodes, which are used to describe the influence of inspection data on compliance behavior status and the triggering relationship of compliance behavior on regulatory clauses and rectification measures.

[0052] To characterize the cross-layer connection structure, a cross-layer adjacency matrix is ​​defined, where the connection relationship from the data layer to the behavior layer is represented by the data-to-behavior adjacency submatrix, and the connection relationship from the behavior layer to the rule layer is represented by the behavior-to-rule adjacency submatrix.

[0053] A hierarchical causal graph structure is generated by combining the node set of the three-layer subgraph with the set of cross-layer directed edges.

[0054] Optionally, the step of inputting the hierarchical causal graph structure into the hierarchical causal graph neural network, performing causal propagation and node state updates, and obtaining the causal propagation result graph specifically includes:

[0055] The hierarchical causal graph structure is input into the hierarchical causal graph neural network, and feature initialization operations are performed on all nodes in the hierarchical causal graph structure to obtain the initial node feature matrix.

[0056] Each node corresponds to an initial feature vector. The initial feature vectors of all nodes together form the initial node feature matrix, where each vector has the same dimension and belongs to the real number space. The initial features of the nodes are taken from the node feature set in the multi-layer subgraph, covering all nodes in the data layer, behavior layer and rule layer.

[0057] The hierarchical causal graph neural network includes an intra-layer propagation module, a cross-layer propagation module, and a directional attention module;

[0058] In each subgraph, intra-layer propagation is performed. Based on the initialized node feature matrix, the nodes are aggregated and their states are updated. For each node, the corresponding node feature vector is extracted from the neighboring nodes according to the adjacency relationship. These feature vectors are then weighted and summed according to the causal propagation weight coefficient to obtain the aggregated intermediate representation. The intermediate representation is multiplied by the learnable weight matrix and a bias term is added. After processing by a non-linear activation function, the updated feature vector of the node in the current propagation layer is generated.

[0059] A directional attention mechanism with parent-child mask constraints is used to calculate the causal propagation weight coefficient of each directed edge;

[0060] In the cross-layer propagation module, for node pairs with cross-layer directed edges, bottom-up feature transfer and state update operations are performed. For each cross-layer directed edge, its parent node belongs to the upper layer subgraph and its child node belongs to the lower layer subgraph. The current layer feature vector of the parent node, the current layer feature vector of the child node, and the causal propagation weight coefficient are obtained. The feature vector of the child node itself and the feature vector of the parent node modulated by the causal propagation weight coefficient are multiplied by the corresponding cross-layer propagation weight matrix, and a bias term is added. After processing by a non-linear activation function, the updated feature vector of the child node in the next layer is generated.

[0061] After completing intra-layer and cross-layer propagation, the above process is iterated until the number of propagations reaches the preset layer threshold, and all nodes have completed multiple rounds of feature updates, forming the final set of node states.

[0062] The feature vectors of each node in the last layer of propagation are merged in order to form a causal propagation result matrix. The causal propagation result matrix records the final semantic state of the nodes in the entire hierarchical causal graph structure. The causal propagation result matrix is ​​combined with the hierarchical causal graph structure to generate a causal propagation result graph containing a set of nodes, a set of edges, and the final node feature matrix.

[0063] Optionally, the step of performing causal inference based on the causal propagation result graph to determine the violation status, risk level, and triggering clauses of the target facility, and generating a causal inference result set specifically includes:

[0064] From the causal propagation result diagram, locate the set of nodes associated with the target emission facility, identify the set of behavioral layer nodes and the set of rule layer nodes, and extract the final node feature vectors corresponding to the violation judgment nodes and risk classification nodes for the set of behavioral layer nodes.

[0065] For the rule layer nodes, the final feature vector of the clause node is extracted to obtain the feature vector of the violation judgment node, the feature vector of the risk classification node, and the feature vector of the clause node.

[0066] Based on the feature vector of the violation judgment node, the violation probability of the target emission facility is calculated. The feature vector of the violation judgment node is multiplied with the violation readout weight vector and a bias term is added to obtain a linear combination result. The Sigmoid function is applied to the linear combination result for nonlinear mapping, and the violation probability value between zero and one is output, thus obtaining the violation probability of the target emission facility.

[0067] Based on the risk classification node feature vector, calculate the probability distribution of the risk level category of the target emission facility, multiply the risk classification node feature vector with the risk classification readout weight matrix, and add the bias vector to obtain the risk score vector.

[0068] The risk score vector is normalized by applying the Softmax function, which converts it into a probability distribution vector for each risk level. Each dimension corresponds to the probability value of a risk level. The risk grading readout weight matrix and bias vector are automatically learned by the backpropagation algorithm during the training phase of the hierarchical causal graph neural network. The risk level with the highest probability value is used as the risk level determination result.

[0069] The trigger score of the corresponding clause is calculated based on the feature vector of the clause node. The feature vector of each clause node is multiplied by the clause reading weight vector and a bias term is added to obtain the trigger score value of the clause node. The trigger scores of all clause nodes are compared with the preset clause trigger threshold. When the trigger score of the clause node is greater than or equal to the threshold, the clause is determined to be triggered. The triggered clause index set is obtained by filtering and forming the trigger clause set.

[0070] The probability of violation, the risk level assessment results, and the set of triggers for the clauses are combined to form a causal reasoning result set.

[0071] Optionally, the process of receiving a natural language question input by the user, semantically encoding the natural language question, and performing semantic matching and path retrieval with the causal inference result set to generate a structured answer specifically includes:

[0072] The system receives a natural language question input by the user and feeds the natural language question as input text into a semantic encoder for encoding processing. The semantic encoder is a text representation network built based on a pre-trained language model.

[0073] The semantic encoder performs word segmentation, embedding, and context modeling operations on the question text, and outputs a question semantic vector to represent the semantic information of the natural language question;

[0074] Extract candidate node sets and candidate path starting point sets from the causal propagation result graph and causal inference result set;

[0075] The candidate node set consists of all nodes in the hierarchical causal graph structure, and the corresponding final node feature vector constitutes the candidate node feature set. The candidate path starting point set is generated based on the key nodes in the causal reasoning result set, including violation judgment nodes, risk classification nodes, and triggered clause nodes.

[0076] Based on the semantic vector of the question and the feature set of candidate nodes, the semantic relevance score of the semantic vector of the question and the feature vector of candidate nodes is calculated. According to the preset screening principle, the set of question-related nodes is determined by either the relevance threshold or by selecting the top few nodes with the highest relevance.

[0077] In the hierarchical causal graph structure, the set of problem-related nodes is used as anchor points for path retrieval. Starting from each anchor node, the adjacent nodes that can be reached are traversed sequentially along the causal direction to generate a candidate path sequence consisting of multiple nodes and directed edges.

[0078] For each candidate path, the generated path score is obtained based on the semantic relevance score of each node and the causal propagation weight coefficient between adjacent nodes.

[0079] Combining the violation probability and risk level judgment results with the set of triggering clauses in the causal reasoning result set, the candidate path set is sorted according to the path scoring results, and the target causal chain path and node status with the highest score are selected.

[0080] A structured response is generated based on the target causal chain path and the causal reasoning result set. The structured response includes emission data nodes related to the question, violation judgment results, risk level results, and triggered regulatory clauses and corresponding rectification suggestions.

[0081] Optionally, the output structured answer content, and the generation of a visual causal explanation result based on the causal chain path, showing the complete reasoning path, specifically includes:

[0082] The system outputs structured answers and generates a visual causal explanation graph based on the set of nodes and their directed edge connections in the causal chain path. The multimodal causal nodes in the causal chain path are distributed to three levels—data layer, behavior layer, and rule layer—according to their hierarchical position in the hierarchical causal graph structure.

[0083] The data layer displays image feature nodes, numerical feature nodes, and text feature nodes related to emission indicators; the behavior layer displays violation judgment nodes and risk level nodes related to violation status judgment; and the rule layer displays regulatory clause nodes and rectification suggestion nodes related to compliance clause judgment.

[0084] Based on the causal dependencies between nodes, directed edges are drawn from the data layer to the behavior layer and then to the rule layer to form a causal path graph;

[0085] Based on the final feature vector state of each node and the context information within the path, the node status is dynamically labeled in the graph, including emission values, violation labels, risk level results, and the triggered regulatory clause numbers.

[0086] The generated causal path map is displayed as a visual causal explanation result, presenting a complete causal reasoning path.

[0087] The beneficial effects of this invention are:

[0088] This invention introduces a hierarchical causal graph neural network to achieve end-to-end modeling of environmental inspection data, from multimodal feature fusion to causal semantic reasoning. This effectively overcomes the problems of information fragmentation and insufficient reasoning ability in compliance analysis found in traditional rule-based or semantic matching question-and-answer systems. At the data level, the system can simultaneously process three types of heterogeneous inputs: images, numerical data, and text, forming a unified multimodal causal node representation, providing a complete data semantic foundation for subsequent causal modeling. At the structural level, the hierarchical causal graph structure constructed in this invention, comprising a data layer, a behavior layer, and a rule layer, achieves dynamic association between emission data, violations, and regulatory clauses through cross-layer causal dependency connections. This gives causal relationships directionality and propagation capability, thereby generating interpretable causal paths during the reasoning process.

[0089] At the algorithmic level, this invention employs a hierarchical causal graph neural network for causal propagation and node state updates. It utilizes a directional attention mechanism to learn the causal weights between nodes, enabling the effective transmission and reinforcement of causal information between parent and child nodes, thereby improving the accuracy of violation determination and risk identification. At the application level, the system uses a semantic matching module to associate natural language questions with causal reasoning results, achieving a semantic closed loop from causal reasoning to intelligent question answering. A causal chain visualization module displays the reasoning path for emission indicators, violation status, regulatory clauses, and rectification suggestions. This invention automates, interprets, and intelligentizes the environmental compliance question-and-answer process, improving the reliability of compliance analysis and decision support capabilities. Attached Figure Description

[0090] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0091] Figure 1 This is a flowchart of a method for an environmental compliance intelligent question-answering system based on causal graph reasoning proposed in this invention.

[0092] Figure 2 This is a schematic diagram of a hierarchical causal graph structure in an environmental compliance intelligent question-answering system based on causal graph reasoning proposed in this invention.

[0093] Figure 3 This is a schematic diagram of a hierarchical causal graph neural network structure in an environmental compliance intelligent question-answering system based on causal graph reasoning proposed in this invention.

[0094] Figure 4 This is a schematic diagram of the structure of an environmental compliance intelligent question-answering system based on causal graph reasoning proposed in this invention. Detailed Implementation

[0095] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0096] refer to Figures 1-4 An intelligent question-answering system for environmental compliance based on causal graph reasoning, comprising:

[0097] The natural language input module is used to receive natural language questions input by the user;

[0098] The multimodal data acquisition module is used to collect environmental inspection data of the emission facilities to be inspected, and obtain the raw inspection dataset;

[0099] The multimodal feature encoding module performs multimodal feature extraction processing on the original inspection dataset to generate multimodal causal node representations;

[0100] The causal graph construction module is used to construct a hierarchical causal graph structure based on the multimodal causal node representation and to make cross-layer connections through causal dependency edges;

[0101] The causal graph neural reasoning module is used to input the hierarchical causal graph structure into the hierarchical causal graph neural network, perform node state updates and causal propagation based on the directional attention mechanism, and generate a causal propagation result graph.

[0102] The causal state reasoning module is used to perform causal reasoning operations based on the causal propagation result diagram, determine the non-compliance status, risk level and triggering clauses of the target facility, and generate a causal reasoning result set;

[0103] The semantic matching and question answering module is used to semantically encode natural language questions and perform semantic matching and path retrieval with the causal inference result set to generate structured answer content.

[0104] The causal chain visualization module is used to display the causal reasoning graph based on the causal paths obtained through matching;

[0105] The compliance question-and-answer output module is used to output structured question-and-answer results and corresponding causal chain graphs.

[0106] In this embodiment, the modules are interconnected using the following method:

[0107] Collect environmental inspection data of the emission facilities to be tested, obtain additional attribute information associated with the inspection task, and obtain the original inspection dataset;

[0108] Multimodal feature extraction is performed on the original inspection dataset to obtain multimodal causal node representations;

[0109] Based on multimodal causal node representation, a data layer subgraph, a behavior layer subgraph, and a rule layer subgraph are established. Cross-layer directed connections are established between the subgraphs according to causal dependencies to construct a hierarchical causal graph structure.

[0110] The hierarchical causal graph structure is input into the hierarchical causal graph neural network, and causal propagation and node state update are performed to obtain the causal propagation result graph.

[0111] Based on the causal propagation result graph, perform causal reasoning operations to determine the non-compliance status, risk level, and triggering clauses of the target facility, and generate a causal reasoning result set;

[0112] It receives natural language questions input by users, performs semantic encoding on the natural language questions, performs semantic matching and path retrieval with the causal inference result set, and generates structured answer content;

[0113] It outputs structured answers and generates visual causal explanations based on causal chain paths, showcasing the complete reasoning path.

[0114] In this embodiment, the process of collecting environmental inspection data from the emission facilities to be inspected, obtaining additional attribute information associated with the inspection task, and obtaining the original inspection dataset specifically includes:

[0115] Environmental inspection data of the target emission facility is collected by scanning terminals or mobile inspection equipment configured at the emission facility site. The environmental inspection data includes image data, numerical detection data and text description data.

[0116] The image data consists of images of the facility's appearance, operating status, and emission outlet taken by the inspection personnel. The numerical detection data includes temperature, pressure, flow rate, pH value, COD, ammonia nitrogen concentration, and suspended solids content. The text description data includes operating status descriptions, maintenance records, and fault notes filled in by the inspection personnel.

[0117] Obtain additional attribute information associated with the inspection task, including facility type identifier, inspection timestamp, geographic location information, and responsible person number;

[0118] Image data, numerical detection data, text description data, and attribute information are uniformly encapsulated to generate the original inspection dataset.

[0119] In this embodiment, the step of performing multimodal feature extraction processing on the original inspection dataset to obtain multimodal causal node representations specifically includes:

[0120] Image data from the original inspection dataset is input into a multi-scale convolutional coding network. The multi-scale convolutional coding network includes three parallel branches, which use convolutional layers with kernel sizes of 3×3, 5×5, and 7×7 to perform feature extraction, respectively, to obtain image feature vectors at three scales. The feature vectors at the three scales are then concatenated and batch normalized to obtain the image feature vector.

[0121] Numerical detection data is input into a numerical coding network to perform embedding mapping and nonlinear feature transformation. Each dimension of numerical data is embedded into a high-dimensional space through linear mapping and nonlinear activation function to obtain an embedding vector. All embedding vectors are concatenated to form a coding matrix. The coding matrix is ​​subjected to two fully connected layers and nonlinear transformation to extract global numerical features and obtain numerical feature vectors.

[0122] The text description data is input into the pre-trained language model encoder to perform context modeling on the text, calculate the semantic embedding of each word, input the semantic embedding of all words into the bidirectional Transformer structure, perform self-attention calculation and semantic aggregation, and obtain the text feature vector.

[0123] Image feature vectors, numerical feature vectors, and text feature vectors are input into a multimodal gating fusion module. The feature vectors of the three modalities are concatenated to form a joint feature representation. The joint feature representation is then input into a weight generation network to calculate the gating weight coefficients corresponding to each modality. The gating weight coefficients are used to represent the importance distribution of each modality in the current sample.

[0124] The original feature vector of each modality is multiplied by the corresponding gating weight coefficient to form a weighted modal feature vector. The three weighted modal feature vectors are summed element by element to obtain the fused output feature vector, which serves as the representation of the multimodal causal node.

[0125] In this embodiment, the step of establishing a data layer subgraph, a behavior layer subgraph, and a rule layer subgraph based on multimodal causal node representation, and establishing cross-layer directed connections between the subgraphs according to causal dependencies to construct a hierarchical causal graph structure specifically includes:

[0126] A data layer subgraph is constructed based on the multimodal causal node representation vector. A data layer node set is defined based on the feature variables related to emission indicators, facility status, inspection time and geographical location in the multimodal causal node representation. The data layer node set includes flow rate nodes, temperature nodes, pH value nodes, chemical oxygen demand nodes, facility type nodes, geographical location nodes and time nodes.

[0127] Assign corresponding node feature vectors to each data layer node, and use linear transformation weight matrix and bias term to map the multimodal causal node representation to obtain the feature representation of each data layer node;

[0128] A set of directed edges in the data layer is established based on the statistical dependencies and structural priors between data layer nodes. The directed edges in the data layer represent the causal directional relationships between nodes and constitute the set of data layer edges.

[0129] A data layer adjacency matrix is ​​generated based on the data layer edge set to represent the directed connection relationships between data layer nodes, forming a data layer subgraph containing node feature sets and adjacency relationships;

[0130] Construct a behavior layer subgraph and define a behavior layer node set based on the feature variables related to corporate emission compliance behavior in the multimodal causal node representation vector. The behavior layer node set includes violation judgment nodes, risk classification nodes, and exceedance level nodes.

[0131] Assign a corresponding node feature vector to each behavior layer node, and use a linear transformation weight matrix and bias term to map the multimodal causal node representation to obtain the feature representation of each behavior layer node;

[0132] The set of directed edges between behavioral layer nodes is determined based on historical violation records and risk assessment data. These directed edges represent the causal dependencies between violation status, risk level, and exceedance level.

[0133] A behavior layer adjacency matrix is ​​generated based on the behavior layer edge set to describe the directed connection relationships between behavior layer nodes, forming a behavior layer subgraph containing the node feature set and adjacency relationships;

[0134] Construct a rule layer subgraph and define a set of rule layer nodes based on the feature variables related to environmental regulations knowledge in the multimodal causal node representation vector. The rule layer nodes include regulatory clause nodes, penalty standard nodes, and rectification suggestion nodes.

[0135] Assign corresponding node feature vectors to each rule layer node, and use linear transformation weight matrix and bias term to map multimodal causal node representation vectors to obtain feature representation of each rule layer node;

[0136] A set of directed edges at the rule layer is established based on the citation relationships between legal provisions, the relationship between penalty standards and the application of provisions, and the relationship between rectification suggestions and the constraints of provisions. The directed edges at the rule layer are used to describe the causal dependencies between legal provisions, penalty standards, and rectification suggestions.

[0137] A rule-layer adjacency matrix is ​​generated based on the rule-layer edge set to represent the directed connection structure between rule-layer nodes, forming a rule-layer subgraph containing the node feature set and adjacency relationship;

[0138] Establish a set of cross-layer directed edges. Based on the semantic relationships between nodes in the data layer subgraph, behavior layer subgraph, and rule layer subgraph, establish sets of cross-layer directed edges between the data layer and behavior layer, and between the behavior layer and rule layer, respectively.

[0139] The cross-layer connection edges include directed edges from data layer nodes to behavior layer nodes and directed edges from behavior layer nodes to rule layer nodes, which are used to describe the influence of inspection data on compliance behavior status and the triggering relationship of compliance behavior on regulatory clauses and rectification measures.

[0140] To characterize the cross-layer connection structure, a cross-layer adjacency matrix is ​​defined, where the connection relationship from the data layer to the behavior layer is represented by the data-to-behavior adjacency submatrix, and the connection relationship from the behavior layer to the rule layer is represented by the behavior-to-rule adjacency submatrix.

[0141] By combining the node set of the three-layer subgraph with the cross-layer directed connection edge set, a hierarchical causal graph structure is generated to support the complete merging of the hierarchical causal graph structure and the directional propagation of causal information.

[0142] In this embodiment, the step of inputting the hierarchical causal graph structure into the hierarchical causal graph neural network, performing causal propagation and node state updates, and obtaining the causal propagation result graph specifically includes:

[0143] The hierarchical causal graph structure is input into the hierarchical causal graph neural network, and feature initialization operations are performed on all nodes in the hierarchical causal graph structure to obtain the initial node feature matrix.

[0144] Each node corresponds to an initial feature vector, which represents the semantic and attribute information of the node in the current level. The initial feature vectors of all nodes together form the initial node feature matrix, where each vector has the same dimension and belongs to the real number space. The initial features of the nodes are taken from the node feature set in the aforementioned multi-layer subgraph, covering all nodes in the data layer, behavior layer and rule layer.

[0145] The hierarchical causal graph neural network includes an intra-layer propagation module, a cross-layer propagation module, and a directional attention module. The intra-layer propagation module is used to perform node feature aggregation and updating within each layer subgraph. The cross-layer propagation module is used to realize information transmission from parent layer nodes to child layer nodes along cross-layer directed edges. The directional attention module is used to calculate the causal propagation weights of each directed edge and adjust the information flow intensity.

[0146] In each subgraph, intra-layer propagation is performed. Based on the initialized node feature matrix, the nodes are aggregated and their states are updated. For each node, the corresponding node feature vector is extracted from the neighboring nodes according to the adjacency relationship. These feature vectors are then weighted and summed according to the causal propagation weight coefficient to obtain the aggregated intermediate representation. The intermediate representation is multiplied by the learnable weight matrix and a bias term is added. After processing by a non-linear activation function, the updated feature vector of the node in the current propagation layer is generated.

[0147] The causal propagation weight coefficients for each directed edge are calculated using a directional attention mechanism with parent-child mask constraints.

[0148] ;

[0149] in, The causal propagation weight coefficients for directed edges in layer l are... An element of a parent-child mask matrix is ​​a given element if and only if a directed edge exists. The value is 1 if the condition is met, and 0 otherwise, to constrain the direction of causal propagation. Let be a set of directed edges, representing causal dependencies between nodes, including intra-layer edges and cross-layer edges. This is the parent node in the hierarchical cause-effect graph structure. For child nodes in a hierarchical cause-effect graph structure. The set of adjacent nodes of the parent node. Index for the parent node, Index the target node (child node). Index all adjacent nodes of the parent node. For the parent node at the 1st Feature vectors in layer propagation, For the first The layer can learn a linear transformation weight matrix. For the directional attention parameter vector, This is a vector concatenation operator. For directed edges Directional encoding or causal strength parameter, For linear rectified activation functions with leakage coefficients, It is an exponential function used for attention weight normalization calculation;

[0150] This invention improves the accuracy and interpretability of causal information propagation by introducing a directional attention mechanism with parent-child mask constraints. This mechanism achieves adaptive modeling of the causal influence strength between different nodes by calculating causal propagation weight coefficients for each directed edge. In this calculation, the attention weights are obtained by exponentially weighting the linear transformation result of the node feature vectors through a linear activation function with leakage correction, and then normalized to achieve a probabilistic distribution, thereby highlighting key propagation directions among multiple candidate causal paths. The exponential term in the formula is used to enhance the weight of highly relevant causal edges, while the introduction of the LeakyReLU function avoids the gradient vanishing problem when features are sparse. By limiting the propagation direction through the parent-child mask matrix, this invention can effectively suppress the interference of invalid dependency edges in hierarchical causal graph neural networks. Experimental results show that after adopting this mechanism, the system outperforms traditional models in compliance judgment accuracy, causal chain consistency rate, and question-answering response time, proving that this mechanism achieves efficient, stable, and interpretable causal inference in compliant environments.

[0151] In the cross-layer propagation module, for node pairs with cross-layer directed edges, bottom-up feature transfer and state update operations are performed. For each cross-layer directed edge, its parent node belongs to the upper-layer subgraph and its child node belongs to the lower-layer subgraph. The current-layer feature vector of the parent node, the current-layer feature vector of the child node, and the causal propagation weight coefficient are obtained. The feature vector of the child node itself and the feature vector of the parent node modulated by the causal propagation weight coefficient are multiplied by the corresponding cross-layer propagation weight matrix, and a bias term is added. After processing by a non-linear activation function, the updated feature vector of the child node in the next layer is generated. Through this process, the causal information transfer from the parent node to the lower node is realized, forming a node feature update result containing cross-layer information flow.

[0152] After completing intra-layer and cross-layer propagation, the above process is iterated until the number of propagations reaches the preset layer threshold, and all nodes have completed multiple rounds of feature updates, forming the final set of node states.

[0153] The feature vectors of each node in the last layer of propagation are merged in order to form a causal propagation result matrix. The causal propagation result matrix records the final semantic state of the nodes in the entire hierarchical causal graph structure. The causal propagation result matrix is ​​combined with the hierarchical causal graph structure to generate a causal propagation result graph containing a set of nodes, a set of edges, and the final node feature matrix.

[0154] In this embodiment, the step of performing causal inference based on the causal propagation result graph to determine the violation status, risk level, and triggering clauses of the target facility, and generating a causal inference result set specifically includes:

[0155] From the causal propagation result diagram, locate the set of nodes associated with the target emission facility, identify the set of behavioral layer nodes and the set of rule layer nodes. For the set of behavioral layer nodes, extract the final node feature vectors corresponding to the violation judgment nodes and risk classification nodes for subsequent violation judgment and risk classification. For the rule layer nodes, extract the final feature vectors of the clause nodes for subsequent regulatory clause trigger judgment, and obtain the feature vectors of violation judgment nodes, risk classification nodes, and clause nodes.

[0156] Based on the feature vector of the violation judgment node, the violation probability of the target emission facility is calculated. The feature vector of the violation judgment node is multiplied by the violation readout weight vector, and a bias term is added to obtain a linear combination result. The linear combination result is then subjected to a Sigmoid function for nonlinear mapping, and the violation probability value between zero and one is output. The violation readout weight vector and the bias term are model parameters automatically learned by the backpropagation algorithm during the training stage of the hierarchical causal graph neural network. Its training objective function uses the real violation labels in the historical inspection samples as the supervision signal. The parameters are updated by minimizing the cross-entropy loss between the predicted probability and the real label. After iterative optimization, the violation readout weight vector and the bias term can represent the nonlinear mapping relationship between violation features and compliance status. Finally, the violation probability result of the target emission facility is obtained, which is used to represent the confidence level of its current compliance status.

[0157] Based on the risk classification node feature vector, calculate the probability distribution of the risk level category of the target emission facility, multiply the risk classification node feature vector with the risk classification readout weight matrix, and add the bias vector to obtain the risk score vector.

[0158] The risk score vector is normalized by applying the Softmax function, which converts it into a probability distribution vector for each risk level. Each dimension corresponds to the probability value of a risk level. The risk grading readout weight matrix and bias vector are automatically learned by the backpropagation algorithm during the training phase of the hierarchical causal graph neural network. The risk level with the highest probability value is used as the risk level determination result.

[0159] The trigger score of the clause is calculated based on the feature vector of the clause node. The feature vector of each clause node is multiplied by the clause readout weight vector and a bias term is added to obtain the trigger score value of the clause node. The trigger scores of all clause nodes are compared with a preset clause trigger threshold. When the trigger score of the clause node is greater than or equal to the threshold, the clause is determined to be triggered. The triggered clause index set is obtained by filtering and forming the trigger clause set. The clause readout weight vector and the bias term are both learned by backpropagation algorithm during the training stage of the hierarchical causal graph neural network.

[0160] The probability of violation, the risk level assessment results, and the set of triggers for the clauses are combined to form a causal reasoning result set.

[0161] In this embodiment, receiving the natural language question input by the user, semantically encoding the natural language question, and performing semantic matching and path retrieval with the causal inference result set to generate structured answer content specifically includes:

[0162] The system receives a natural language question input by the user and feeds the natural language question as input text into a semantic encoder for encoding processing. The semantic encoder is a text representation network built based on a pre-trained language model, which is used to extract the contextual semantic features of the input text.

[0163] The semantic encoder performs word segmentation, embedding, and context modeling operations on the question text, and outputs a question semantic vector to represent the semantic information of the natural language question;

[0164] Extract candidate node sets and candidate path starting point sets from the causal propagation result graph and causal inference result set;

[0165] The candidate node set consists of all nodes in the hierarchical causal graph structure, and the corresponding final node feature vectors constitute the candidate node feature set for semantic matching calculation. The candidate path starting point set is generated based on the key nodes in the causal reasoning result set, including violation judgment nodes, risk classification nodes and triggered clause nodes, which are used as the starting point for path retrieval.

[0166] Based on the question semantic vector and the candidate node feature set, the semantic relevance score between the question semantic vector and the candidate node feature vector is calculated. Specifically, the cosine similarity between the question semantic vector and the final node feature vector of each candidate node is calculated to obtain the semantic relevance value of each node. The cosine similarity is calculated by dividing the inner product of the question semantic vector and the node feature vector by the product of the magnitudes of the two vectors. The result is between zero and one and is used to represent the semantic closeness between the two. According to the preset screening principle, the set of question-related nodes is determined by either the relevance threshold or by selecting the top few nodes with the highest relevance.

[0167] In the hierarchical causal graph structure, the set of problem-related nodes is used as anchor points for path retrieval. Starting from each anchor node, the adjacent nodes that can be reached are traversed sequentially along the causal direction to generate a candidate path sequence consisting of multiple nodes and directed edges.

[0168] For each candidate path, the semantic relevance score of each node on the path and the causal propagation weight coefficient between adjacent nodes are summed to obtain the path score.

[0169] In the calculation of path score, semantic relevance is used to measure the degree of semantic matching between a node and the input question, and causal propagation weight coefficient is used to reflect the strength of causal relationship between nodes. The two are weighted and summed to obtain the path score value. Finally, a path score result is generated for each candidate path.

[0170] Combining the violation probability and risk level judgment results with the set of triggering clauses in the causal reasoning result set, the candidate path set is sorted according to the path scoring results, and the target causal chain path and node status with the highest score are selected.

[0171] A structured response is generated based on the target causal chain path and the causal reasoning result set. The structured response includes emission data nodes related to the question, violation judgment results, risk level results, and triggered regulatory clauses and corresponding rectification suggestions.

[0172] In this embodiment, the output of structured answer content and the generation of visualized causal explanation results based on the causal chain path, demonstrating the complete reasoning path, specifically includes:

[0173] The system outputs structured answers and generates a visual causal explanation graph based on the set of nodes and their directed edge connections in the causal chain path. The multimodal causal nodes in the causal chain path are distributed to three levels—data layer, behavior layer, and rule layer—according to their hierarchical position in the hierarchical causal graph structure.

[0174] The data layer displays image feature nodes, numerical feature nodes, and text feature nodes related to emission indicators; the behavior layer displays violation judgment nodes and risk level nodes related to violation status judgment; and the rule layer displays regulatory clause nodes and rectification suggestion nodes related to compliance clause judgment.

[0175] Based on the causal dependencies between nodes, directed edges are drawn from the data layer to the behavior layer and then to the rule layer to form a causal path graph;

[0176] Based on the final feature vector state of each node and the context information within the path, the node status is dynamically labeled in the graph, including emission values, violation labels, risk level results, and the triggered regulatory clause numbers.

[0177] The generated causal path map is displayed as a visual causal explanation result, presenting a complete causal reasoning path.

[0178] Example 1:

[0179] To verify the feasibility of this invention in practice, it was applied to the intelligent question-and-answer system for emission monitoring of a large manufacturing enterprise. This enterprise has multiple production lines and emission facilities, requiring daily real-time monitoring and compliance assessment of waste gas, wastewater, and noise emissions. Traditional inspection systems rely on manual rule matching and fixed knowledge bases, failing to effectively integrate image, numerical, and textual data. This results in disconnects between multimodal information, low accuracy in violation determination, and difficulty in providing clear causal explanations.

[0180] In actual deployment, the system runs in the enterprise's environmental monitoring data center, connected to on-site monitoring equipment, mobile inspection terminals, and text reporting systems. The multimodal data acquisition module automatically aggregates approximately 1200 inspection records daily, including about 300GB of image data, about 50,000 numerical detection data entries, and about 2000 text description data entries. The multimodal feature encoding module extracts features and performs gating fusion on the three types of data, generating multimodal causal node representations, with an average processing time of approximately 0.9 seconds per group. Subsequently, the hierarchical causal graph construction module establishes a causal graph structure with data, behavior, and rule layers based on emission indicators, behavioral states, and regulatory knowledge, generating approximately 32,000 nodes and 86,000 causal edges. After causal propagation and node state updates are performed by the hierarchical causal graph neural network, the system forms a causal propagation result graph, used for subsequent violation identification and compliance Q&A.

[0181] During the application phase, users can input questions in natural language, such as "Why were the emissions from the exhaust gas monitoring station deemed non-compliant?" The system automatically parses the semantics, retrieves relevant causal chains, and outputs a structured answer containing emission data nodes, violation status nodes, and regulatory clause nodes. The system's identification results show that the nitrogen oxide concentration at this emission point exceeds the standard by 32.5%, showing a strong correlation with historical emission data characteristics. Simultaneously, the system automatically matches the corresponding compliance clause category based on regulatory knowledge and generates targeted rectification suggestions. The average response time for the entire reasoning and question-answering process is 2.1 seconds, and users can view the complete causal reasoning path in a visual interface.

[0182] During a six-month trial run, the system achieved a 96.8% consistency rate between automatically identified violations and manual review results, improved the accuracy of question-and-answer semantic matching by approximately 20 percentage points, and increased the compliance judgment accuracy to over 95%. The overall response time was reduced by approximately 50% compared to traditional systems. This invention demonstrates significant fusion efficiency and reasoning accuracy under complex environmental data, realizing a shift in environmental compliance analysis from "experience-based judgment" to "causal reasoning," and possesses scalability and high practical value.

[0183] Table 1 Performance Comparison Results of Environmental Compliance Intelligent Question Answering Systems

[0184] ;

[0185] As shown in Table 1 above, the environmental compliance intelligent question-answering system based on causal graph reasoning proposed in this invention significantly improves upon traditional systems in several key performance indicators. First, in terms of compliance judgment accuracy, the system achieves 95.4%, an improvement of approximately 15 percentage points compared to traditional rule-matching systems. This indicates that by introducing a hierarchical causal graph neural network and a multimodal feature fusion mechanism, the system can more fully capture the deep causal relationships between emission data, equipment status, and text descriptions, thereby achieving high-precision compliance judgments for complex environmental events. Second, in terms of violation identification recall, the system achieves 94.1%, an improvement of nearly 13 percentage points compared to traditional systems, effectively reducing missed detections and demonstrating its stronger sensitivity and robustness to edge anomalies in the causal propagation and node state update stages.

[0186] In terms of question-answering performance, the semantic matching accuracy of the system of this invention reaches 91.8%, which is about 18 percentage points higher than that of traditional systems. This result reflects that the semantic encoding and causal path matching mechanism adopted by the system can more accurately understand user questions and retrieve relevant causal chain paths based on causal inference results, thereby generating answers with logical consistency and explanatory power. In terms of response speed, the average question-answering response time of the system is shortened to 2.1 seconds, which is more than 50% less than that of traditional methods, indicating that the hierarchical causal graph structure and attention propagation mechanism have a significant effect on improving inference efficiency.

[0187] Furthermore, the consistency rate of causal inference results in this invention reaches 96.8%, achieving a high degree of consistency between question-and-answer results and manual review results, demonstrating the interpretability and reliability of the model. In terms of data processing capabilities, this invention's system processes an average of 1200 inspection data entries per day, 2.3 times that of traditional systems, and the average processing latency per group is reduced from 1.8 seconds to 0.9 seconds, showcasing excellent engineering application performance and scalability. In summary, the above results indicate that this invention achieves overall performance improvements in multimodal data fusion, causal inference accuracy, question-and-answer semantic understanding, and response efficiency, providing a highly accurate, highly interpretable, and highly real-time intelligent solution for environmental compliance management.

[0188] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An environment compliance intelligent question answering system based on causal diagram reasoning, characterized in that, The method comprises the following steps: a natural language input module for receiving a user input natural language question; a multi-modal data acquisition module for collecting environmental inspection data of the emission facility to be detected to obtain an inspection original data set; a multi-modal feature encoding module for performing multi-modal feature extraction processing on the inspection original data set to generate a multi-modal causal node representation; a causal graph construction module for constructing a hierarchical causal graph structure according to the multi-modal causal node representation, and connecting across layers through causal dependency edges; a causal graph neural reasoning module for inputting the hierarchical causal graph structure into a hierarchical causal graph neural network, performing node state updating and causal propagation based on a directional attention mechanism, and generating a causal propagation result graph; a causal state reasoning module for performing causal reasoning operations based on the causal propagation result graph to determine the violation state, risk level and trigger clause of the target facility, and generating a causal reasoning result set; a semantic matching and question answering module for performing semantic encoding on the natural language question, and performing semantic matching and path retrieval with the causal reasoning result set to generate structured answer content; a causal chain visualization module for displaying a causal reasoning graph based on the matched causal path; a compliance question answering output module for outputting structured question answering results and corresponding causal chain graphs; The causal graph neural reasoning module is implemented in the following way: Based on the multi-modal causal node representation, a data layer subgraph, a behavior layer subgraph and a rule layer subgraph are established, and cross-layer directed connection edges are established between the subgraphs according to the causal dependency relationship to construct a hierarchical causal graph structure; specifically including: Based on the multi-modal causal node representation vector, a data layer subgraph is constructed, and a data layer node set is defined according to the multi-modal causal node representation; Each data layer node is assigned a corresponding node feature vector, and a multi-modal causal node representation is mapped using a linear transformation weight matrix and a bias term to obtain a feature representation of each data layer node; According to the statistical dependency and structural prior between the data layer nodes, a data layer directed edge set is established, a data layer adjacency matrix is generated according to the data layer edge set, and a data layer subgraph is formed; A behavior layer subgraph is constructed, a behavior layer node set is defined according to the multi-modal causal node representation, and each behavior layer node is assigned a corresponding node feature vector. The multi-modal causal node representation is mapped using a linear transformation weight matrix and a bias term to obtain a feature representation of each behavior layer node. According to the historical violation records and risk assessment data, a behavior layer directed edge set between the behavior layer nodes is determined, a behavior layer adjacency matrix is generated according to the behavior layer edge set, and a behavior layer subgraph is formed; A rule layer subgraph is constructed, a rule layer node set is defined according to the multi-modal causal node representation, and each rule layer node is assigned a corresponding node feature vector. The multi-modal causal node representation vector is mapped using a linear transformation weight matrix and a bias term to obtain a feature representation of each rule layer node. According to the reference relationship between the regulations and clauses, the penalty standard and clause application relationship, and the rectification suggestion and clause constraint relationship, a rule layer directed edge set is established, a rule layer adjacency matrix is generated according to the rule layer edge set, and a rule layer subgraph containing a node feature set and an adjacency relationship is formed. A cross-layer directed connection edge set is established, and according to the semantic association relationship of the nodes in the data layer subgraph, the behavior layer subgraph and the rule layer subgraph, the cross-layer directed connection edge set is established between the data layer and the behavior layer and between the behavior layer and the rule layer, and a cross-layer adjacency matrix is defined; The hierarchical causal graph structure is generated by combining the node set of the three-layer subgraph with the cross-layer directed connection edge set.

2. The environment compliance intelligent question answering system based on causal diagram reasoning according to claim 1, characterized in that, The modules are realized by the following methods: Collecting environmental inspection data of the to-be-detected emission facility, obtaining additional attribute information associated with the inspection task, and obtaining an inspection original data set; Performing multi-modal feature extraction processing on the inspection original data set to obtain multi-modal causal node representations; Based on the multi-modal causal node representations, a data layer subgraph, a behavior layer subgraph and a rule layer subgraph are established, cross-layer directed connection edges are established between the subgraphs according to the causal dependency relationship, and a hierarchical causal graph structure is constructed; Inputting the hierarchical causal graph structure into the hierarchical causal graph neural network to perform causal propagation and node state updating to obtain a causal propagation result graph; Based on the causal propagation result graph, a causal reasoning operation is performed to determine the violation state, risk level and trigger clause of the target facility, and a causal reasoning result set is generated; Receiving a natural language question input by a user, performing semantic encoding on the natural language question, and performing semantic matching and path retrieval on the causal reasoning result set to generate structured answer content; Outputting the structured answer content and generating a visual causal explanation result based on the causal chain path to display the complete reasoning path.

3. The environment compliance intelligent question answering system based on causal diagram reasoning according to claim 2, characterized in that, The environmental inspection data includes image data, numerical detection data and text description data.

4. The environment compliance intelligent question answering system based on causal diagram reasoning according to claim 2, characterized in that, The multi-modal feature extraction processing on the inspection original data set to obtain multi-modal causal node representations specifically includes: Inputting the image data in the inspection original data set into a multi-scale convolutional coding network to perform multi-scale feature extraction on the image information to obtain an image feature vector; Inputting the numerical detection data into a numerical coding network to perform embedding mapping and nonlinear transformation on each detection parameter to obtain a numerical feature vector; Inputting the text description data into a pre-trained language model encoder to perform context semantic modeling on the text content to output a text feature vector; Inputting the image feature vector, the numerical feature vector and the text feature vector as inputs, performing fusion processing through a multi-modal gating fusion module to obtain multi-modal causal node representations.

5. The environment compliance intelligent question answering system based on causal diagram reasoning according to claim 2, characterized in that, The inputting of the hierarchical causal graph structure into the hierarchical causal graph neural network to perform causal propagation and node state updating to obtain a causal propagation result graph specifically includes: Inputting the hierarchical causal graph structure into the hierarchical causal graph neural network to perform feature initialization operation on all nodes in the hierarchical causal graph structure to obtain an initialized node feature matrix; The hierarchical causal graph neural network includes an intra-layer propagation module, a cross-layer propagation module and a directional attention module; Performing intra-layer propagation operation in each layer subgraph, performing feature aggregation and state updating on the nodes according to the initialized node feature matrix to generate an updated feature vector of the nodes in the current propagation layer; The directional attention mechanism with parent-child mask constraint is used to calculate the causal propagation weight coefficient of each directed edge; In the cross-layer propagation module, for each cross-layer directed edge, the current layer feature vector of the parent node, the current layer feature vector of the child node, and the causal propagation weight coefficient are obtained, the bottom-up feature transmission and state update operation is performed, and the updated feature vector of the child node in the next layer is generated; After completing the intra-layer propagation and cross-layer propagation, the iteration process is continued until the propagation times reach the preset layer threshold, all nodes have completed multiple rounds of feature updating, and the final node state set is formed; The feature vectors of the nodes in the last layer propagation are merged in order to form a causal propagation result matrix, and the causal propagation result matrix and the hierarchical causal graph structure are combined to generate a causal propagation result graph.

6. The environment compliance intelligent question answering system based on causal diagram reasoning according to claim 2, characterized in that, The causal reasoning operation based on the causal propagation result graph is performed to determine the violation state, risk level and trigger clause of the target facility, and a causal reasoning result set is generated, which specifically includes: From the causal propagation result graph, the node set associated with the target emission facility is located, and the violation judgment node feature vector, the risk classification node feature vector and the clause node feature vector are obtained; Based on the violation judgment node feature vector, the violation probability of the target emission facility is calculated; Based on the risk classification node feature vector, the risk level category probability distribution of the target emission facility is calculated, and the risk level with the maximum probability value is taken as the risk level determination result; Based on the clause node feature vector, the corresponding clause trigger score is calculated, all clause node trigger scores are compared with the preset clause trigger threshold, and the trigger clause set is selected; The violation probability, risk level determination result and clause trigger set are combined to form the causal reasoning result set.

7. The environment compliance intelligent question answering system based on causal diagram reasoning according to claim 2, characterized in that, The natural language question input by the user is received, the natural language question is semantically encoded, and the semantic matching and path retrieval are performed with the causal reasoning result set to generate structured answer content, which specifically includes: The natural language question input by the user is received, and the natural language question is taken as an input text and sent to a semantic encoder for encoding processing to output a question semantic vector; From the causal propagation result graph and the causal reasoning result set, a candidate node set and a candidate path starting point set are extracted; Based on the question semantic vector and the candidate node feature set, the semantic relevance score of the question semantic vector and the candidate node feature vector is calculated, and the question related node set is determined according to the preset screening principle; On the hierarchical causal graph structure, the path retrieval is performed with the question related node set as the anchor point to generate a candidate path sequence; For each candidate path, based on the semantic relevance score of each node and the causal propagation weight coefficient between adjacent nodes, a path score result is obtained; The candidate path set is sorted according to the path score result in combination with the violation probability, risk level determination result and trigger clause set in the causal reasoning result set, and the target causal chain path and node state with the highest score are selected; The structured answer content is output, and the visual causal explanation result based on the causal chain path is generated to display the complete reasoning path, which specifically includes:

8. The environment compliance intelligent question answering system based on causal diagram reasoning according to claim 2, characterized in that, ​ Output structured answer content, generate a visual causal explanation graph based on the node set contained in the causal chain path and the directed edge connection relationship; Distribute the multi-modal causal nodes in the causal chain path according to the hierarchical position in the hierarchical causal graph structure into three levels of data layer, behavior layer and rule layer; According to the causal dependence relationship between nodes, draw directed edges from the data layer to the behavior layer, and then to the rule layer, to form a causal path graph; According to the final feature vector state of each node and the context information in the path, dynamically label the node state in the graph, and display the generated causal path graph as a visual causal explanation result.

Images